Make grading-schema prose_fields extend the default set instead of replacing it

[justinmclean]

Summary

A per-step grading-schema.json replaces the runner's default prose-field set
rather than adding to it, so a prose field added later silently reverts to
exact-match grading. One schema is also now dead config.

Background

Prose fields are graded by a model; everything else is compared exactly. When a
step ships a grading-schema.json naming one prose field, the default set stops
applying to that step. Add a notes or summary field to that step's
expected.json later and it is exact-matched, failing on any wording
difference, with nothing to indicate why.

Where to look

• evals/release-verify-rc/*/fixtures/grading-schema.json - both set
  prose_fields: ["paste_recipe"].
• evals/release-audit-report/step-2-assemble-record/fixtures/grading-schema.json -
  sets prose_fields: ["record_markdown"], but that is no longer a compared
  expected.json key, so the schema acts on nothing.
• tools/skill-evals/src/skill_evals/runner.py - where the default set and the
  per-step override are resolved.

Acceptance criteria

• Document whether a per-step schema replaces or extends the default, in the
  skill-evals README.
• The behaviour matches the documented decision, with a test.
• The dead release-audit-report schema is removed or given a real field.
• uv run --directory tools/skill-evals --group dev pytest passes.

Estimated effort

~2 hours for someone new to the codebase.